Pinned Tweet
1/5 🧵
Excited to share that our paper "Crafting Reversible SFT Behaviors in Large Language Models" has been accepted to #NeurIPS2026! 🚀
Can we design SFT behaviors that are causally localized and completely reversible at inference time?
Paper: arxiv.org/abs/2605.06632
1
9
4/5 🔄 SFT-Eraser: Reversibility via Activation Matching
At inference time, zero weight modifications are needed. We optimize a soft prompt (SFT-Eraser) that targets the Carrier channels via activation matching to instantly "turn off" or reverse the SFT behavior on demand.
1
Super excited to share that two of our papers have been accepted to the Main Conference at #ACL2026! 🎉 Huge thanks to my amazing co-authors and collaborators.
Check out the thread below for a quick TL;DR on our work covering Mechanistic Interpretability and LLM Agents. 👇🧵
1
1
28
📄 1/ Retrieval Heads are Dynamic
We demonstrate that retrieval heads in LLMs vary dynamically across timesteps. Furthermore, the model's hidden state actually encodes a predictive signal for future retrieval patterns.
🔗 arxiv.org/abs/2602.11162
1
19
📄 2/ How Memory Management Impacts LLM Agents
We explored how memory addition/deletion impacts long-term behavior of LLM Agents. And we offer guidance for robust memory design!
🔗 arxiv.org/abs/2505.16067
1
13
Yuping Lin retweeted
Our paper for LLM watermark is accepeted by NAACL findings! We proposed a new method to strengthen the robustness of watermark agains paraphrase using the semantics. This is very meaningful factor for the practical application! Please find the paper at openreview.net/forum?id=hbMR…
Yuping Lin retweeted
🙋♂️What is the wildest dream for graph foundation models?
🎯Graph across domains → a single model → all the downstream
🙋♀️Can we achieve that?
✅Yes! UniAug: Cross-Domain Graph Data Scaling with Diffusion Models
📃arxiv.org/pdf/2406.01899
Yuping Lin retweeted
🚀 Excited to share our latest research on enhancing privacy in RAG systems! arxiv.org/pdf/2406.14773
Our paper introduces SAGE, a novel approach using synthetic data to protect sensitive information while maintaining high utility.
#AI #Privacy #MachineLearning #RAG #DataSecurity
✨ Excited to share our new preprint "Towards Understanding Jailbreak Attacks in LLMs: A Representation Space Analysis"! arxiv.org/abs/2406.10794
🔍 We delve into why some jailbreak attacks succeed by exploring harmful and harmless prompts in the LLM's representation space.
🤔 Understanding why some jailbreak methods succeed helps us build more resilient LLMs. This is crucial for safe AI deployment in real-world applications. #Research #AI #LLM #Jailbreak
1
34
Code available at: github.com/yuplin2333/repres…
32







